Eight years ago, one of the most painful moments in Chinese industry history occurred. In April 2018, Zhongxing Communications suddenly shut down — a giant company with 80,000 employees and revenues exceeding one trillion yuan. Everything happened because America decided to cut supplies. Without Qualcomm chips, without Google’s Android system — the company was completely trapped. This lesson was not forgotten.

Fast forward to 2026. Now the situation is completely different. This time, the war is not just about chips — it’s a much deeper battle called CUDA. You may not have heard of it before, but every AI developer in the world depends on it. CUDA is a platform from Nvidia that controls everything — from training to deployment. And here lies the real shortage. The more the industry relies on CUDA, the greater the lack of an independent alternative.

But Chinese companies have chosen a different path this time. Instead of trying to directly imitate Nvidia, they focused on algorithms. DeepSeek is a perfect example — their model contains 671 billion parameters, but only activates 37 billion during operation. This means significantly lower costs. The model was trained with 2,048 H800 processors over 58 days at a cost of just $5.6 million. Compare that to the estimated $78 million to train GPT-4. The difference is enormous.

The result? DeepSeek is 25 to 75 times cheaper than Claude. And this price changed the market. In only February 2026, the use of Chinese models on OpenRouter increased by 127% in just three weeks. A year ago, their market share was no more than 2%. Now it’s approaching 60%. This is no coincidence — it’s a structural shift.

But the biggest challenge remained: obtaining enough computational power for training. That’s where local chips came in. Loongson 3C6000 and Taichu Yuanqi have already begun handling real training tasks. In January 2026, Zhipu AI launched the GLM-Image model — the first image generation model trained entirely on Chinese-made chips. This is a qualitative shift from simple inference to real training.

Behind all this stands Huawei’s Ascend system. By the end of 2025, the number of developers there exceeded 4 million. Building this independent ecosystem is exactly what Japan did in the 1980s when faced with similar American pressures. Japan chose to be “the best” in a system dominated by others. China chose to build its own system.

Energy also plays a crucial role. The United States is facing a real electricity crisis — data centers now consume 4% of American electricity, and that number will double by 2030. In contrast, China produces 2.5 times more electricity than the US. Industrial electricity prices in western China are about $0.03 per kilowatt-hour — a quarter to a fifth of US prices. This is a huge advantage.

Now DeepSeek serves 37 languages and is expanding globally. 30.7% of users are from China, 13.6% from India, 6.9% from Indonesia. In sanctioned markets, its share ranges between 40% and 60%. 58% of new AI startups in China are now using it.

Three local chip companies published their results on February 27, 2026. Some achieved profits for the first time, while others lost billions. But these losses are not failures — they are investments in building an independent ecosystem. Every dollar lost is a dollar in R&D, software support, and engineer training.

The war for computational power is no longer about “can we survive?” Now the question is “how much will we spend to stay independent?” And the answer is clear — any price. Because independence is no longer an option.